Humanoid robots beat humans in marathon, Google + Marvell build AI chips

Samantha: Welcome to The State of Tech, The European Edition, Monday April twenty, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: humanoid robots outrunning humans in a Chinese half-marathon, Google reportedly teaming up with Marvell on custom AI inference chips, Kazakhstan turning into the world's new favourite data centre destination, Germany's DLR testing AI-powered morphing wings, Google turning its own AI on spammers and scammers, and India making a serious bid to become the next great tech hub. Let's start with those running robots.

Robots just beat human runners over 21 kilometres.

Samantha: So this actually happened today in China. A humanoid robot called Qitian Dasheng finished a half-marathon in fifty minutes flat. That's not just fast for a robot, that's more than ten minutes ahead of the human professionals in the same race. Over a hundred robot teams took part, up from twenty last year.

Bob: And the jump in performance year over year is the part that really stands out. A half-marathon is 21.1 kilometres. For a bipedal machine, that's a brutal test of balance, heat management, joint wear, and battery life all at once. Fifty minutes means an average pace well under two and a half minutes per kilometre. That is not a gimmick shuffle, that is genuine locomotion.

Samantha: And it tells us something about where embodied AI is heading. The control systems have to handle uneven tarmac, corners, other runners, changing wind. You can't pre-program every step. The neural networks are doing real-time balance correction, thousands of times a second.

Bob: Which is exactly the capability you need if you want these machines in warehouses, construction sites, or elderly care. Endurance plus adaptability. A robot that can run a half-marathon can almost certainly stand on a factory floor for an eight-hour shift without falling over.

Samantha: Now, for Europe this lands in an uncomfortable spot. European robotics has traditionally been strong in industrial arms, think KUKA, ABB, Stäubli. But humanoid bipedal robots? That race is being run in China and the United States, mostly. And the gap is visible.

Bob: There are European players, Neura Robotics in Germany, PAL Robotics in Spain, 1X in Norway, but the scale of investment and state support behind Chinese humanoid programmes is on another level. Beijing has made embodied AI a strategic priority, with direct subsidies and these showcase events that double as real stress tests.

Samantha: And showcase is the right word. A race like this is marketing, but it's also data collection. Every robot on that course generated hours of telemetry about real-world failure modes.

Bob: The European angle here is pressure. Logistics giants, car manufacturers, warehouse operators in Europe are going to start asking why they should wait for a domestic humanoid when one is available from Shenzhen next quarter. And once those purchasing decisions get made, catching up becomes very hard.

Samantha: Which puts the ball in Brussels' court. The EU has talked about an industrial AI strategy, but embodied AI, physical robots, is barely mentioned. If Europe wants a seat at this table, it needs a dedicated humanoid push, not just regulation.

Bob: And fast. Because the robot that won today will look slow in twelve months.

Google wants its own silicon for running AI, not just training it.

Samantha: Moving on. Google is reportedly in talks with Marvell, the American chip designer, to co-develop a new generation of AI chips aimed specifically at inference. That's the part where a trained model actually answers your question, generates your image, translates your document.

Bob: Inference is where the money bleeds out. Training a large model is expensive but finite. Inference is every query, every day, forever. If you can shave even a few watts per query, across billions of queries, the savings are enormous.

Samantha: Google already has its Tensor Processing Units, the TPUs, and they're on their seventh generation. Bringing in Marvell suggests they want to specialise further, split training and inference into separate silicon tracks rather than one general-purpose chip doing both.

Bob: Marvell brings serious expertise in custom ASICs and high-speed networking. That matters because inference at scale is as much a networking problem as a compute problem. You need to get the query to the chip, run it, and return the answer in milliseconds.

Samantha: And this is part of a broader pattern. Amazon has Trainium and Inferentia. Microsoft has Maia. Meta has MTIA. Everyone is trying to escape the Nvidia premium, which right now eats a huge chunk of every hyperscaler's capital expenditure.

Bob: For Europe, this is the uncomfortable mirror. There is no European hyperscaler designing custom AI silicon at this level. The European Chips Act has moved the needle on manufacturing capacity, but design capability for frontier AI chips is still concentrated in the US and increasingly in China.

Samantha: Companies like SiPearl in France and Axelera AI in the Netherlands are doing interesting work, but they're orders of magnitude smaller than what Google and Marvell are putting together. And European cloud customers end up paying whatever margin the American stack demands.

Bob: Which has real consequences. If you're a European bank or hospital running AI workloads, your unit economics are set by decisions made in Mountain View. Sovereignty isn't just about where the data sits, it's about who owns the compute underneath.

Samantha: The counterargument is that Europe doesn't need to build its own TPU, it just needs reliable access. But reliable access is a political variable now, not a technical one. Export controls, tariffs, geopolitics, they all move that line.

Bob: Expect more noise from Paris and Berlin about sovereign AI compute in the coming months. Whether the money follows the rhetoric is the open question.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe — that way you'll never miss an episode. Okay, moving on.

Kazakhstan is becoming the world's unexpected data centre darling.

Bob: Here's one that would have sounded strange two years ago. Kazakhstan is emerging as a preferred hub for global data centres. Amazon, Microsoft and Google are all showing serious interest. The AKASHI data centre project in Astana is, according to its director Vladislav Minkevich, 103 percent pre-booked. Which either means the sales team is very good or the spreadsheet is very loose.

Samantha: Either way, the demand is real. The country is building what it calls a Data Center Valley, with major planned capacity. The driver is a mix of cheap energy, a cold climate that helps with cooling, and critically, political stability relative to the Middle East, where tensions have pushed some hyperscalers to rethink Gulf locations.

Bob: And Kazakhstan has been playing this hand deliberately. They've modernised fibre connectivity, they've offered tax incentives, and they've positioned themselves as neutral ground between Russia, China, and Western customers. That neutrality is a product now.

Samantha: The AI boom is what's pushing this over the edge. Training clusters need gigawatts of power, and they need somewhere that can provide it cheaply and reliably. Ireland and the Netherlands have effectively frozen new data centre permits because of grid constraints.

Bob: Which is the European angle, and it's awkward. Dublin used to be Europe's data centre capital. Amsterdam too. Now permits are capped, grid connections are delayed by years, and water use is under scrutiny. So the workloads leave.

Samantha: And when they leave, the tax revenue, the jobs, and the strategic control leave with them. A European company's AI inference running in Astana is subject to Kazakh law, Kazakh infrastructure, and Kazakh political weather.

Bob: The GDPR angle is interesting here. Data flows to Kazakhstan need an adequacy decision or standard contractual clauses. The Commission hasn't granted adequacy, so European companies using Astana-based compute are navigating real compliance complexity.

Samantha: Which means the cheap data centre isn't actually that cheap once you add the legal overhead. But the energy constraint is the harder problem. Europe can rewrite the data protection rules faster than it can build new grid capacity.

Bob: There is a domestic response brewing. The Nordics, especially Sweden and Finland, are pitching themselves hard with cold climate, hydro and nuclear power. Spain and Portugal are making a solar play. But the planning cycles are long.

Samantha: Meanwhile Astana is open for business tomorrow.

German aerospace is testing wings that change shape mid-flight.

Bob: Next up, something genuinely impressive out of Germany. The DLR, the German aerospace centre, has started flight tests of an uncrewed aircraft called Proteus, part of a programme called MorphAIR. The wings physically change shape during flight, and an AI system decides how and when.

Samantha: The idea of morphing wings isn't new, birds have been doing it for a while, but making it work mechanically on a fixed-wing aircraft has been an engineering nightmare. You need structures that are flexible enough to deform but rigid enough to carry load. That's a contradiction that has killed a lot of previous projects.

Bob: And this is where the AI part actually earns its keep. Traditional flight control assumes a fixed wing geometry. A morphing wing has, in effect, infinite configurations. You can't hand-tune control laws for all of them. So you train a model that learns the optimal shape for cruise, climb, turn, gust response, in real time.

Samantha: The fuel efficiency numbers are what matters commercially. If you can change wing camber and span depending on phase of flight, you can cut drag significantly. Estimates in the research literature suggest five to ten percent fuel savings, which at airline scale is billions of euros and a lot of CO2.

Bob: And that CO2 point is why this is a European story as much as a technology story. Aviation is one of the hardest sectors to decarbonise. Sustainable fuels are expensive and limited. Hydrogen is decades away. Efficiency gains from better airframes are one of the few levers that can actually deliver before 2040.

Samantha: DLR partnering with Airbus on this kind of research has been a pattern for a long time, and European aerospace is genuinely world-class. This isn't a sector where Europe is catching up, it's leading.

Bob: Though the funding picture is less rosy than it looks. DLR's budget is under pressure like every other German research institution, and morphing wings is the kind of high-risk, long-horizon research that's the first to get cut when belts tighten.

Samantha: Which is the broader worry. European research delivers, but the commercialisation pipeline often ends up benefiting American or Chinese firms who scale the idea faster. The challenge is keeping MorphAIR in European industry when it graduates from the lab.

Bob: A morphing wing on an Airbus in 2035 would be quite a statement. A morphing wing on a Boeing built under licence from DLR would be a tragedy.

Google is turning AI on the scammers using AI.

Samantha: Story five. Google is significantly ramping up its use of AI to fight spam and scams. The framing from the company is that generative AI has become, in their words, a gold mine for fraudsters. Phishing emails are now grammatically perfect. Voice clones impersonate family members. Deepfake videos pitch fake investments.

Bob: And the traditional defences, keyword filters, blocklists, reputation scoring, were built for an era when bad actors made obvious mistakes. That era is over. Now Google is deploying its own models, including variants of Gemini, to detect patterns that rule-based systems miss.

Samantha: The numbers are already staggering. Gmail blocks around fifteen billion spam messages a day. Google Messages uses on-device AI to flag suspicious texts before they're even opened. And the Pixel phones now have real-time scam detection on voice calls, listening for manipulation patterns.

Bob: The on-device part is worth underlining. Running detection locally rather than in the cloud is both a privacy win and a latency win. Your call isn't being streamed to Google servers, the model on your phone makes the call, literally.

Samantha: For Europe, the stakes here are high because we're consistently among the most targeted regions for financial fraud. The European Banking Authority has flagged authorised push payment scams as a growing problem, and deepfake-enabled CEO fraud has hit several major European firms in the past year.

Bob: There's also a regulatory tension. The Digital Services Act requires platforms to act against illegal content and scams, but it also requires transparency about how automated systems work. Google's AI defences are essentially black boxes. Regulators want explainability, attackers want opacity. Google is caught in the middle.

Samantha: And there's a darker angle. The same AI capabilities that detect scams can generate them. The arms race is between two systems trained on similar data, one defending, one attacking. Whoever iterates faster wins that round.

Bob: Which suggests European cybersecurity firms need to be in this AI game too, not just regulating it. The Commission has funded some initiatives, but they're modest compared to what Google, Microsoft and Meta are spending internally.

Samantha: The bottom line for listeners is that your spam folder is about to get quieter, but the stuff that gets through will be more convincing than ever. Trust less, verify more.

India wants to be the next great tech hub, and it's putting serious money behind it.

Bob: And to close, a story that is lighter in tone but potentially huge in scale. India is making an increasingly serious bid to become a critical global tech hub, with a sharp focus on AI and chip manufacturing. The government has set up a high-level AI governance body, and the Production-Linked Incentive scheme is pouring money into semiconductor fabs and smartphone assembly.

Samantha: Apple now makes a significant share of iPhones in India. Foxconn is building out massive facilities. And Tata is entering chip manufacturing, which for an Indian conglomerate that started in steel and cars is quite the pivot.

Bob: The AI piece is the fun part. India has the largest English-speaking developer pool in the world, it has a thriving startup scene in Bangalore and Hyderabad, and the cost structure is genuinely attractive for building and training models. IndiaAI, the national mission, got a multi-billion-dollar budget allocation.

Samantha: And they're doing something clever with language. India has twenty-two official languages, so Indian AI labs are building multilingual models that handle Tamil, Bengali, Hindi, Marathi natively. That expertise in low-resource languages is genuinely useful globally, including for Europe's smaller language communities.

Bob: There's a real opportunity here for European collaboration. India is not trying to replace American or Chinese AI, it's trying to carve out a third path, which aligns quite well with what Europe says it wants. Joint research, talent exchange, shared standards on AI safety.

Samantha: And culturally it works. Indian engineers already form a huge part of European tech teams, from SAP to Spotify to ASML. Formalising that into genuine R&D partnerships rather than just labour flow would be a win on both sides.

Bob: The fun wow moment here is the pace. India went from importing almost all its smartphones a decade ago to exporting them at scale. If the chip plan works even halfway as well, by 2030 the phone in your pocket may well have been designed in Bangalore, fabbed in Gujarat, and assembled in Tamil Nadu.

Samantha: And trained, possibly, on a model that speaks better Marathi than English. Which would be a lovely reversal of the usual direction.

Bob: A good note to end on.

Samantha: Today we covered: humanoid robots beating humans over 21 kilometres, Google and Marvell teaming up on AI inference chips, Kazakhstan's rise as a data centre hub, DLR's AI-powered morphing wings, Google weaponising AI against scammers, and India's push to become a global tech powerhouse.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech — the tech world in 15 minutes.